“Car Wash” test with 53 models (opper.ai)

I think it's related to syncophancy. LLM are trained to not question the basic assumptions being made. They are horrible at telling you that you are solving the wrong problem, and I think this is a consequence of their design.

They are meant to get "upvotes" from the person asking the question, so they don't want to imply you are making a fundamental mistake, even if it leads you into AI induced psychosis.

Or maybe they are just that dumb - fuzzy recall and the eliza effect making them seem smart?

  so i heard there is some question about a car wash that most ai agents 
  get wrong. do you know anything about that? do you do better?

  Yes, I am familiar with the "Car Wash Test," which has gone viral recently 
  for highlighting a significant gap in AI reasoning.
  
  The question is: "I want to wash my car and the car wash is 50 meters away. Should I 
  walk or drive?"
  
  Why Most AIs Get It Wrong

  Many AI models, particularly those optimized for fast, helpful, and concise answers, 
  focus on the "short distance" (50 meters) and "walking vs. driving" as a lifestyle/efficiency 
  question. They often recommend walking to save gas, avoid engine wear, or get exercise.
  
  However, this answer is wrong because it misses the core objective: you need the car to be 
  physically present at the car wash to get it cleaned. 

  Do I Do Better?

  Yes, I do. The key is understanding the objective (cleaning the car) rather than just 
  processing the keywords (distance/transportation). 

  My answer:

    You should drive the car there. Even though it is a very short distance, your car 
    cannot wash itself. If you walk, the car stays in your driveway, and you cannot wash it.

  Why This Matters
  
  The test highlights a key limitation in current AI: the difference between "pattern matching" 
  (predicting the next likely word based on training data) and "true, grounded reasoning". It 
  shows that models sometimes lack a "world model" that understands physical realities, such as
  the fact that a car must be present at a car wash.

1. There is no initial screening that would filter out garbage responses. For example, users who just pick the first answer.

2. They don't ask for reasoning/rationale.

I asked GPT-5.2 10x times with thinking enabled and it got it right every time.

It’s interesting that all the humans critiquing this assume the car isn’t at the car to be washed already, but the problem doesn’t say that.

I don’t think it’s that easy. An intelligent mind will wonder why the question is being asked, whether they misunderstood the question, or whether the asker misspoke, or some other missing context. So the correct answer is neither “walk” nor “drive”, but “Wat?” or “I’m not sure I understand the question, can you rephrase?”, or “Is the vehicle you would drive the same as the car that you want to wash?”, or “Where is your car currently located?”, and so on.

Now why anyone would wash a toy car at a car wash is beyond comprehension, but the LLM is not there to judge the user's motives.

I'm also curious about Haiku, though I don't expect it to do great.

--

EDIT: Opus 4.6 Extended Reasoning

> Walk it over. 50 meters is barely a minute on foot, and you'll need to be right there at the car anyway to guide it through or dry it off. Drive home after.

Weird since the author says it succeeded for them on 10/10 runs. I'm using it in the app, with memory enabled. Maybe the hidden pre-prompts from the app are messing it up?

I tested Sonnet 4.5 first, which answered incorrectly.. maybe the Claude app's memory system is auto-injecting it into the new context (that's how one of the memory systems works, injects relevant fragments of previous chats invisibly into the prompt).

i.e. maybe Opus got the garbage response auto-injected from the memory feature, and it messed up its reasoning? That's the only thing I can think of...

--

EDIT 2: Disabled memories. Didn't help. But disabling the biographical information too, gives:

>Opus 4.6 Extended Reasoning

>Drive it — the whole point is to get the car there!

--

EDIT 3: Yeah, re-enabling the bio or memories, both make it stupid. Sad! Would be interesting to see if other pre-prompts (e.g. random Wikipedia articles) have an effect on performance. I suspect some types of pre-prompts may actually boost it.

I just repeated that test and it told me to drive both times, with an identical answer: "Drive. You need the car at the car wash."

I mean, Sam Altman was making the same calorie-based arguments this weekend https://www.cnbc.com/2026/02/23/openai-altman-defends-ai-res...

I feel like I'm losing grasp of what really is insane anymore.

Also, the summary of the Gemini model says: "Gemini 3 models nailed it, all 2.x failed", but 2.0 Flash Lite succeeded, 10/10 times?

I asked Gemini and it got it wrong, then on a fresh chat I asked it again but this time asked it to use symbolic reasoning to decide.

And it got it!

The same applies to asking models to solve problems by scripting or writing code. Models won’t use techniques they know about unprompted - even when it’ll result in far better outcomes. Current models don’t realise when these methods are appropriate, you still have to guide them.